t0mas88

Realistically a 32GB card is what you need to run Gemma4:31b well. Doing it in 24GB leaves almost no room for context / KV cache.

Agree, I run QFN on my AI Pro R9700 (32GB) using https://github.com/Niko1221/Strata It's even faster than on a Strix Halo. Typically I get around 90 tokens/sec with a small context, down to 70 tokens/sec when it's near the maximum context usage.

I'm still not sure whether Donald himself is just dumb or whether he knows he can get away with saying things because his voters are dumb. Neither option looks good in an international context...

Not only the Norway thing, but he also (pretends?) doesn't understand how NATO works. Being angry about NATO members, a defensive treaty, not joining an offensive campaign in Iran.

The system where every CA gets to sign certificates for every possible domain makes no sense anymore. There are too many CAs and each one of them could compromise the whole system.

Why aren't SSL certificates tied to something like DNSSEC where a CA is responsible only for one tld? Since domain ownership is enough to get a certificate signed anyway, that wouldn't change security properties.

Commercial support in Linux seems to be loosing its advantage over Debian.

Ubuntu (including Ubuntu Pro) still doesn't have a CVE-2026-99999 patch while Debian does have it. For commercial operators a KVM guest escape vulnerability is quite important.

Same here. I run Qwen3.8 Flash Next on my local machine and combine it with Opus 5.5 for plan making. The Qwen model feels like it's around Opus 4.x level performance most of the time.

But once in a while it gets into long reasoning loops, which is annoying. It seems less with the Swift model but I haven't benchmarked the difference between that one and the original.

Interesting idea. I guess the main challenge is to create some kind of harness to make the input volume appropriate for an LLM to evaluate. Billions of the same log events from a raw stream will be cost prohibitive and not find you much interesting output. But a trend in some value changing would.

As a human I look at Datadog or Grafana plus some log searches, maybe those need to become tools for the agent.

Ah that explains why the content is great but the writing style in English is AI.

Nice write up. I've been optimising build duration and feedback speed across a few projects recently. Agentic coding is putting more emphasis on best practices that we all dragged our feet on for many years :-)

But... Tell the AI to write more concise and remove typical AI-isms next time. The post is twice as long as it needs to be and keeps doing the "The slow linter isn't the problem" fluffy Claude thing.

Not spilling coffee during a roll is a solved problem: https://youtu.be/V9pvG_ZSnCc

Interesting, that explains why getting 4x48 = 192gb is around 2400 euros, so 1.2x the price for 1.5x the capacity.

Speed has indeed decreased for the modules you can actually buy, while historically you would get faster and faster modules at lower prices in the years after the introduction is a new DDR standard.

Currently 128GB of normal non-ECC DDR5 as a 2x64 set is around 2000 euro and increasing...

For daily office use there is much more benefit in having more memory than having faster memory. Because apps all started using more memory. Especially anything Google Chrome related.

If you have some unused extra memory it will be used as disk cache, which also helps in the responsiveness of a desktop. Waiting for the disk is orders of magnitude slower than reading from memory cache.

There are, but I've seen several of those price increases in a row so you may be moving a lot. And mail deliverability to Outlook and Gmail from those hosts isn't always great.

I hope so.

The plugin is for AppHaven to allow Claude to deploy containers and databases, check their build status and logs etc. So it can take all of the devops work for AI coded applications out of your hands.

Here is an example of what it does: https://apphaven.eu/#how-it-works

The portal feels very vibe-coded.

I have a Claude plugin on there, but I now see it twice, once with an empty version number and once with 1.0.0 while the latest is actually the 1.1.0 version. Then when I click the 1.0.0 listing under versions the 1.1.0 version appears. Feels like no QA happened on this thing :-)

It is still one step better than the last six weeks, since nothing from their pipeline went live at all during that time: https://github.com/anthropics/claude-plugins-community/issue...

I think we're going to see a huge number of small internal tools that simply wouldn't have been worth building before coding agents. Even more so in SMB, because they have less budget.

The part I don't think you can leave to the agent is everything around the application: backups, authentication/access control and vulnerability scanning. Those things aren't rocket science, but getting them wrong can have a much bigger impact than a bug in the application itself.

That's what I'm building with https://AppHaven.eu My bet is that vibe-coded apps will become cheap and almost disposable, while the platform underneath them needs to provide those boring safety guarantees automatically.

Nice idea, but this is not how flight planning works. The optimization algorithms that airlines use consider a broader range of variables than just fuel, since you also have cost per hour that the engine is running, cost per hour for the crew, and overflight fees for different airspaces.

And the output still has to be a valid flight plan route. You can't just file a flight plan for "I want this zig zag line at this altitude".

Commercial flight planning software does this optimization as well, but it only selects routes that are valid. For example you have to be at a certain altitude to go one way on a route and another altitude to go the other way. ATC can't do their job if you have planes pointing at each other in opposite directions doing 1000+ km/h.

I'm working on a hosting platform for AI-written apps: https://apphaven.eu/

The idea is to make it easier for not-so-technical users to ask their Claude Code or Codex (or similar) agent to deploy the app. And then the platform handles backups, login/access control, vulnerability scanning etc.

A Claude Code plugin is under review at Anthropic to make it very simple to use. For now you can also install it from Github: https://docs.apphaven.eu/agent-integration

That's not really how it works.

The industry does not treat a go around like an error. You'll only ever get questions when you don't go around while procedures clearly require it, like what seems to have happened in this case.

On autopilot use: The pilot is always responsible for the approach. You can definitely have an unstable approach on autopilot if you configure it wrong or ask it to do things that can't be done within the stable approach criteria. A common example is trying to slow down and descend at the same time, most jets can't really do that.

But "responsible" does not necessarily mean to blame in aviation. The industry as a whole follows the Just Culture principles (see: https://www.caa.co.uk/general-aviation/general-aviation-at-t... ) The purpose of a report is to learn from it and improve safety, not to blame the crew.

When building boring web applications with a sizeable team that need to run for a long time. Hiring developers is easy since there are many, there is nearly no magic and the language is quite strict and type safe so it works well with a large team.

And that "team" nowadays may also consist of many AI agents. In my experience Claude Code for example works very well with a typed, slightly boring language with lots of framework and library support. Because it doesn't compile when you get something wrong, instead of getting a vague runtime issue that Claude can't always see.

No, a go around is not a pilot error and it's not punished.

It's almost the opposite. There is automatic monitoring for stable approach criteria (simplified: you must be at speed X with flaps selected etc by altitude Y). So this example would cause an automatic report from the aircraft to the airline, because they were too fast (not within stable approach criteria) and still continued the approach instead of going around.

However even in those cases it's not meant for punishment but for safety improvement. In most professional operations these reports are made anonymous, and only for example someone from a union committee gets to know who the crew were and could have a talk with them if necessary.

AI slop to the max, what a bad article.

Now on the issue itself: Why doesn't Tesla report the version and whether it was FSD or Autopilot?

There is also a baby products online store in the Netherlands that for religious reasons is closed on Sunday. You can browse the site but you can't buy anything.

The page is in Dutch, it explains that they're closed because get believe Sunday is a day for rest: https://www.babypark.nl/zondag/info-bp.html

Their goal isn't only safe storage, it's also being able to sell when needed.

You can't quickly trade with the gold that's in their own safe storage facility in Zeist, because you would need to physically transport it. The reserve at the Bank of England is put there because there is a large market for gold trading there. So you can use the gold immediately to influence the markets.

Putting it in a safe place in another European country wouldn't have added value over storage in Zeist.

Selling to the US is equally complex, every state has their own sales tax rules and whether they apply to your type of product. Some exclude SaaS, some don't. And then for B2B you also have to apply the tax unless the business presents some kind of exemption certificate.

EU VAT within the EU for B2B is easy, make an API call to validate the VAT ID of the customer and you're done.

The visualisation is nice to compare with different engine types as well: https://random.ma.ttias.be/boxermotor

My only complaint would be that there is no inline-6 while that does have some interesting properties in terms of balance compared to an inline-4 or V6.

If you have the right authorisation on a business account you can "direct debit" from any account. It let's you charge a bank account in a similar way as a US creditcard can be charged by number and expiry date. Technically the contract with your bank says you need to have a signed authorisation from the account owner to charge their account, but you don't have to show it to the bank before charging.

The money you get that way isn't guaranteed, the account owner could do a charge back. And if that happens you have to show the signed authorisation from the account owner. If you don't have it, expect your bank to terminate the direct debit ability.

They will have moat in the satellite launching business, which is not useful in the AI datacenter market.

You can put AI chips in datacenters in the desert for far less than $100/kg. With lots of solar power available, the option to easily access your hardware and far less radiation issues.

The datacenter in space story really only exists to make it possible for Musk to sell X to SpaceX and make more money from the IPO. That's all. There is no engineering reason.

I've created a simple project for that, except that by default they open PRs on Forgejo/Github/Gitlab. It's open source and the workflows are config files, so you can configure it to allow you to pull from them instead of creating a PR. Source: https://github.com/smithy-ai/smithy-ai

One way I use it that is similar to what I think you want is to run a local Forgejo instance (just like OP) and have the Dockerised Claude instances work in that, then pull from there to my local git repo, review, test etc and push from my local repo to a remote like Github.